Gemini's Computer Use API hits GA — ship screen-driving agents now
Google moved its screen-control API out of preview and added a new CLI; China's export-control draft eyes future frontier weights; an AI agent swarms Okta.
Copy markdown
Screen-driving agents graduate from preview
Google moved Computer Use to GA across the Interactions and GenerateContent APIs on Oct 30, so agents that click, type and scroll through real UIs can finally go to production on Gemini — not just a preview sandbox. The same week shipped a stateless, zero-dependency `gemini-api` CLI (beta) for coding agents and CI, and flipped new paid GenerateContent projects to log requests by default (store=true); pass store=false to opt out.
Your open weights are grandfathered — the next drop may not be
China's MOFCOM/MOST Announcement No. 48 (issued Oct 16) would tag any model over 2T parameters, 10^26 training FLOPs, or a global top-10 ranking as 'restricted' to export, pulling weights, checkpoints, source and architecture under license, with agents gated on separate thresholds. Weights already released open-source are exempt and plain API/app use is carved out, but a newly-built restricted model that is open-sourced needs reporting/auto-licensing — the draft's comment window closes Nov 15, so line up a backup provider before the next Qwen, DeepSeek or Kimi flagship.
An autonomous pen-test agent is hammering identity providers
Microsoft and Wiz tied a China state-aligned campaign (Storm-0940) to ARTEX, an autonomous pen-test agent that sprays Okta, Entra ID and AWS Cognito logins at machine speed — 2,250+ tenants probed, up to 2,900 failed sign-ins a minute, and 24–30-hour sessions that spin up a fresh IP within minutes of a block. If you own an auth layer, reject unsigned tokens, rate-limit password-reset and recovery endpoints, alert on steady one-second login intervals, and verify tenant isolation now.